Using Interconnected ML Models to Tackle Retail Challenges
summary
Shopify builds a multitask hierarchical classifier (5500+ taxonomy nodes) that uses multilingual BERT for text and MobileNetV2 for images, caches embeddings to avoid recomputing, and backfills across billions of products and tens of billions of images. Those outputs feed Shopify Audiences, which publishes anonymized audience sets to ad platforms and reports improved ad spend efficiency. A separate attribute pipeline uses domain taxonomies and annotators. Infrastructure runs on Kubernetes with Merlin + Ray workspaces, model registry and feature store so models move to production quickly.
AI-written summary · about
chapters (8)
- 0:00Shopify ML journey
- 3:12E-commerce challenges
- 7:25Product categorization automation
- 14:09Multitask product classification
- 25:25Shopify Audiences
- 30:20Visual attributes extraction
- 35:17ML infrastructure at Shopify
- 39:09Q&A
related talks
51mKeynote Interview
37mLeveraging Generative AI for Enhanced E-commerce
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling
22mFrontiers of Machine Learning from the Lens of Retail
44mCorrida do ouro moderna
23mDeep Recommendations for e-Commerce
21mBuilding Scalable ML Systems: Strategies and Insights
38mE-Commerce Product Categorization at Scale